February 26, 2026

VM agents in seconds, orchestration race heats up & more

VM agents in seconds, orchestration race heats up & more

Today’s Overview

Enterprises are accelerating AI adoption through customizable agent platforms, multi-model orchestration, and dedicated hardware that promise higher productivity and reliability. Recent launches from Anthropic, Perplexity, Google, and hardware innovators such as Taalas and MatX illustrate a shift toward integrated, scalable AI services that can be tightly controlled and efficiently executed at scale.

  • Anthropic expands Cowork with department-specific Claude agents and private plugin stores, enabling enterprise-wide AI deployment.
  • Perplexity launches Computer, a multi-model orchestration platform that creates specialized sub-agents across 19 models for autonomous workflows.
  • Google introduces the WebMCP standard, allowing AI agents to invoke structured website actions and improving reliability for SaaS and e-commerce interactions.
  • Taalas releases the HC1 chip embedding Meta’s Llama 3.1 8B model in silicon, delivering about 17,000 tokens per second per user and accelerating real-time LLM applications.
  • MatX secures over $500 million to accelerate AI-accelerated chip designs, underscoring growing investment in specialized hardware for enterprise AI.

Top Stories

Anthropic expands Cowork with department-specific Claude agents

Anthropic has added pre‑built Claude agents tailored for ten enterprise functions such as HR, engineering, banking and wealth management. The update introduces connectors for Google Workspace, DocuSign, FactSet and Harvey, as well as partner plugins like Slack by Salesforce. Enterprises can now create private plugin stores and assign custom agents to specific teams with granular administrative controls. A research preview also lets Claude transition seamlessly between Excel and PowerPoint to automate data analysis and presentation creation.

Read Full Article

Perplexity unveils Computer, a multi-model AI agent platform

Perplexity introduced Computer, a platform that can create specialized sub‑agents to perform tasks such as web browsing, code generation and application integration. Each sub‑agent runs in an isolated sandbox and can be assigned any of the nineteen supported models, enabling long‑running workflows that may persist for months. The service uses a consumption‑based pricing model, with a “Max” tier offering a monthly allowance of 10 000 credits and the option to select the model for each job. Perplexity positions this flexibility as a differentiator from competing autonomous‑agent solutions.

Read Full Article

Kilo introduces KiloClaw for one-minute OpenClaw agent deployment

Kilo announced KiloClaw, a service that lets developers launch OpenClaw agents in production in under a minute without complex infrastructure setup. The platform runs the agents on multi‑tenant virtual machines, providing persistent “always‑on” availability along with built‑in monitoring and security features. KiloClaw integrates with the Kilo Gateway, granting access to more than 500 AI models, and includes the PinchBench tool to benchmark and select the optimal model for a given task.

Read Full Article

Research & Analysis

AlphaEvolve uncovers novel multi-agent learning algorithms

The AlphaEvolve project applied an evolutionary coding agent powered by large language models to automatically generate new multi‑agent learning algorithms for imperfect‑information games. The system produced two variants, VAD‑CFR, which adds adaptive discounting, and SHOR‑PSRO, which incorporates a hybrid optimistic meta‑solver. Empirical tests showed that both algorithms converge more quickly and achieve higher performance than established baselines. These results suggest that automated algorithm discovery can meaningfully advance multi‑agent learning research.

Read Source

Taalas unveils HC1 chip delivering 17,000 tokens per second per user

Taalas introduced the HC1 chip, which embeds Meta’s Llama 3.1 8B model directly onto silicon to achieve roughly 17 000 tokens per second for each user. By integrating the model at the hardware level, the design removes traditional memory bottlenecks that limit LLM inference speed. The first HC1 version is already shipping, and subsequent revisions aim to support larger models with higher fidelity. This approach promises a substantial performance increase for real‑time language‑model applications.

Read Source

Data Engineering for Terminal Agents

The paper presents Terminal‑Task‑Gen, a synthetic pipeline that automatically generates terminal‑focused tasks for training autonomous agents. Using this pipeline, the authors created Terminal‑Corpus, a curated collection of terminal interaction data intended to support research on command‑line agents. The dataset includes a diverse set of commands, outputs and error scenarios to enable systematic evaluation of agent performance. By providing both the generation framework and the corpus, the work aims to advance data engineering practices for terminal‑based AI systems.

Read Source

Aletheia tackles FirstProof mathematics challenge autonomously

Aletheia, built on Gemini 3 Deep Think, attempted the FirstProof mathematics challenge without human assistance and correctly solved six of the ten problems. Independent expert reviewers validated the solutions, with the exception of problem 8 where consensus was not reached. The full set of results and the underlying data have been made publicly available for further analysis. This demonstration highlights the system’s capability to tackle advanced mathematical reasoning tasks autonomously.

Read Source

Trending Tools

  • Google introduces WebMCP for agents

    Google released the WebMCP (Web‑Machine‑Callable‑Procedures) standard, which defines structured actions that AI agents can invoke on websites, such as renewing a subscription or booking a flight. By replacing fragile DOM scraping with reliable API‑like calls, WebMCP makes agent interactions more dependable for SaaS and e‑commerce platforms.

  • Foundation Models SDK for Python

    Apple’s Foundation Models SDK for Python provides a Pythonic interface to its on‑device foundation‑model framework, enabling developers to run inference, stream text generation, and apply type‑safe decorators. The repository includes documentation and examples to help integrate these capabilities into applications.

  • Codex Prompting Guide

    OpenAI’s Codex Prompting Guide offers best‑practice recommendations for using the Codex‑tuned API (gpt‑5.2‑codex), suggesting the default “medium” reasoning setting for typical interactive coding and higher settings for complex, long‑running tasks. Following these guidelines helps developers balance performance with cost efficiency.

Quick Hits

Join the AI Recap Newsletter

Get the latest AI news, research insights, and practical implementation guides delivered to your inbox daily.

By subscribing, you agree to our Terms of Service and Privacy Policy.